OGR's end-to-end slate recommender beats baselines by 48.2% NDCG
Kuaishou's A/B test shows 1.12% more effective views from single-pass slate generation.
A team led by Yang Hu, with contributors from Kuaishou and academic institutions, has published a new paper proposing OGR (Once Generated, Ranked), an end-to-end generative framework for slate recommendation. Traditional slate recommendation pipelines separate candidate generation from ranking, limiting optimization to retrieved items and missing joint utility. Generative recommendation with Semantic IDs (SIDs) promises an end-to-end path, but existing SIDs lack recommendation-aware semantics and local collaborative signals, and next-token prediction doesn't align with slate-level goals.
OGR solves this by introducing TUSID, which fuses item-specific semantic and local collaborative information into hierarchical SIDs, and a pipelined position-wise SID decoder that models global preferences and inter-item dependencies. The framework also includes SPA, a reward-guided conservative policy optimization method that aligns generated slates with user preferences beyond simple likelihood imitation. Offline experiments show OGR outperforms representative baselines, with 48.2% and 27.2% relative NDCG@5 improvements on industrial and public datasets. Online A/B testing on Kuaishou further demonstrates a 1.120% improvement in Effective Views, a key engagement metric. The paper is available on arXiv (2608.17613).
- OGR unifies candidate generation and ranking in a single generative pass over ordered item slates.
- TUSID combines semantic and collaborative signals into hierarchical Semantic IDs for better slate-aware representation.
- Online A/B testing on Kuaishou showed a 1.120% lift in Effective Views, confirming real-world impact.
- SPA policy optimization aligns generation with user reward, improving NDCG@5 by 48.2% (industrial) and 27.2% (public dataset).
Why It Matters
End-to-end slate generation can streamline recommender systems, improving engagement metrics with less pipeline complexity for large-scale platforms.